MMM FAQ for Marketers

Frequently asked questions.

MMM - Basics and what it actually is

  1. MMM (Marketing Mix Modelling) is a method for measuring the effect of marketing activities (often media / channel spend) on sales or a similar KPI, thus measuring ROI. Unlike attribution modelling it does not use user-level tracking data, instead it uses statistical learning on time series (daily spends by channel, daily total sales) to quantify what channel drives how much sales.
  2. MMM is especially useful for brands with upper-funnel activities or offline media, offline sales channels or marketplace sales channels or brands with digital tracking issues.
  3. MMM also measures so-called channel saturation which shows how as you invest more into a channel, you get more sales but the incremental growth is slowing down.
  4. MMM is also used for budget optimization – it can help you find the optimal allocation of marketing budget across channels.

Media Mix Modelling generally includes only media channels as factors influencing the target KPI (e.g. sales), while Marketing Mix Modelling includes a broader set of additional non-media factors such as pricing / discounting. So you can think of Media Mix Modelling as a narrower version of Marketing Mix Modelling.

MMM is gaining a lot of traction especially with digital brands because of several factors:

  • Traditional methods for ROI measurement such as attribution modelling rely on accurate user-tracking – this has become much harder / impossible to achieve with privacy changes (both legal and technical on the browser side), changes in iOS policies etc. The result is that attribution models do not have their “fuel” – complete and accurate user tracking data. MMM does not use user level data and thus is fully resistant to these shifts – it is a privacy-first method of measurement.
  • There has been a lot of progress in the last 5 years in MMM that has made it much more automated and accessible. MMM is still hard but thanks to these advances even mid-sized advertisers can now implement MMM as a core component of their marketing measurement.


MMM works at the level of marketing channels – e.g. Facebook Ads or TikTok or Youtube or Google Ads PMax could be these channels. Depending on your marketing/media mix, you may want to divide larger platforms or media types into subcategories – e.g. split FB Ads or Google Ads into several buckets by campaign type (by targeting, objective, bidding type or creative execution etc), or TV by specific networks and so on.

In total it is usually realistic to work with 6-25 channels in total – it depends on how many channels you have, what their shares are, how many data points you have and other factors. Designing a good channel structure (channel grouping) is one of the important phases of MMM implementation and requires some experience with what is ok vs what can cause problems for MMM.

MMM is not a suitable method for determining ROI on very granular level – such as keyword or ad-group.

No, you don’t install any pixel or any other user tracking – MMM works with aggregate data (e.g. total daily sales and daily spends on Google Pmax, FB ASC, Youtube video ads etc).

It identifies patterns in this data using statistical learning and this way it can quantify the impact of each of the explaining variables (such as spend on Google Pmax) on the explained variable (total sales). In a super-simplified way for illustrative purposes think of MMM as investigating “In the 38.week of last year there was an increase in spend on PMax, is it possible to identify some increase in total sales in W38, W39 etc that can be tied to the change?”

You don’t need any pixel or user tracking for this, you can learn this from the daily time series of PMax costs and other channel costs and total sales – MMM does this, of course in a much more sophisticated way but the principle remains – it investigates daily or weekly time series and using statistical learning algorithms it identifies what channel drives what results.

MMM - Process

The initial implementation usually takes a few weeks and consists of

    1. Making sure the MMM project has clear business goals and the right expectations
    2. Data collection and integration
    3. Data quality assessment
    4. Modelling & feedback cycles with the customer 
    5. Final model – interpretation and recommendations
    6. Final model – production deployment for continuous use

It is important to understand that the more you as the advertiser/business owner are involved the better – some vendors offer a “fully automated no-touch” approach but in our experience this typically results in models that may be mathematically ok but just don’t make business sense and nobody uses them afterwards.

The initial implementation usually takes a few weeks – while it is easy to “get some model” within days, it typically takes quite a few iterations (even with the most advanced AI automation) to arrive at a production-level model. So you should expect first useful insights in a matter of weeks.

MMM is best treated as an iterative process – the model should not be a static thing developed once but rather a continuous iterative process – with new calibration points and new data the model is refined and improved. In our experience:

    1. It takes a few weeks to arrive to a useful production level model – ie not just a “first attempt” or a toy model but actually a good enough model that the marketing department can use
    2. Within the next 3-6 months the model can be significantly improved via additional calibrations and other methods

Generally you will need 2-3 years of daily or weekly data for:

    1. Total sales/revenue (or a similar “outcome” KPI that you want to measure) – this can come from your ERP, CRM, e-shop system or data warehouse, or it is possible to use data from GA/GA4 or a similar web analytics tool. This is the “explained” variable (also called dependent or endogenous variable)
    2. Daily or weekly media costs for online and offline channels. Besides cost it is helpful to have additional metrics that represent the volume of the activity – think impressions for online ads, GRPs for TV and so forth. These are the “explaining” variables (also called independent or exogenous variables)
    3. It is OK if some channels run only sometimes and not the whole 2 or 3 years.

It is important that there is some variance in the volume of the channel activity – eg. sometimes you run more FB ads, sometimes less. In most cases this is not an issue but for example having a football league sponsorship for the whole investigated period would not be a good use case for MMM measurement as there is no variance in it.

There are multiple options how to collect and integrate the data to MMM:

    1. Directly via connectors to main platforms like Google Ads, Facebook, TikTok etc. 
    2. Via one-off export of data in CSV, Excel file or similar
    3. If you have the data in your data warehouse or data lake, the MMM provider can also connect directly there (this usually involves some coordination with your BI/data team)

The data needs to be structured into channels (typically 6-25 channels) – the specific channel grouping used for MMM should be both suitable for the modelling process and useful for you as a marketer (i.e. it should reflect how you budget and manage marketing). This is often not easy and there will be tradeoffs to be made – good channel structure is a very important factor in MMM success and is a part where significant previous experience with MMM and marketing is indispensable.

MMM - and attribution, other measurement FWs

From a business perspective both attribution model and MMM try to answer the same (*) question: How much did marketing channel X drive sales?

  • Attribution
      • Relies on user-level tracking, needs complete and accurate data on users and their “touchpoints” (usually clicks, sometimes impressions). It uses a bottom-up approach: from individual users’ journeys it infers about the value of each channel.
      • There are various attribution models ranging from simple ones (position-based) to more advanced ones (Shapley, Markov, LSTM approaches,…)
      • “Direct” – how the attribution treats direct traffic often affects the result more than what algorithm it uses
      • Attribution can see only the online world: online marketing and online sales. Attribution models won’t help you understand how TV impacts your sales or how Youtube impacts your retail store sales.
      • Some online channels often end up underrepresented even with data driven attribution models – very often the case of channels that do not generate any clicks but impact consumers in other ways (video, influencers,…)
      • Can be very granular and thus good for operational decisions in online marketing (daily, weekly).
      • Is “static” over time – once you have it, it is usually not improved/evolved over time
      • Easier implementation
  • MMM
    • Does not use user-level data and does not need tracking
    • Holistic view of demand drivers – both online and offline media, pricing and discounting, market trends and forces, competition activities – all these things affect your results and can be incorporated into MMM (unlike attribution model)
    • Can work with all sales channels – web, app, marketplaces, own retail store network, retail partners
    • Limited granularity – MMM is not for measuring ROI of individual keywords, ad groups or creatives.
    • More useful for strategic use cases – think planning, budgeting, evaluating major campaigns and promotions, understanding true incremental value of major 10-20 channels, marketing budget optimization etc.
    • Implementation is more complex + evolves over time (if you do MMM right you will have a much better model after 6-12 months than after the initial implementation)

not very good at measuring completely new channels or very small channels.

(*) technically there are differences between what attribution and MMM measure (and not just how) and this is a good topic for analysts but from a marketing manager’s position the purpose and the main business question is the same.

MMM and attribution are not mutually exclusive but rather complementary – MMM is better for strategic budgeting decisions (think weekly, monthly, quarterly horizon) and for measuring channels that are typically underrepresented even by data driven attribution models (video, influencers etc) or cannot be measured by online attribution at all (TV advertising, effect of discounting etc). It also benefits from being completely resistant to user tracking issues (as it does not rely on any user-level data). 

Attribution on the other hand is more suitable for operational measurement in online marketing (think daily and weekly horizon). Advanced advertisers use a triangle of measurement methods:

  • MMM,
  • attribution,
  • and tests/experiments (to validate the results of both MMM and attribution and get the so called “ground truth” results)
  • MMM for holistic view of demand drivers

If you are a smaller or starting brand (spending less than 2m USD a year on marketing), using attribution may be completely sufficient for your needs.

MMM - Tips for success

Get serious about incrementality and testing. Get used to doing experiments in marketing regularly. Explain marketing incrementality to senior leadership in your company – CEO, CFO etc. Explain the limitations of attribution models.

Existing experiment results can significantly speed up any MMM program and will almost certainly lead to a faster value and ROI on your MMM project.

Set up a good data collection process for areas like promotions (promotion calendars and plans), main events affecting your business, discounting or major pricing changes etc.

We recommend refreshing data on a weekly basis – it is possible to refresh even daily and for some advertisers it makes sense, but on the whole weekly refreshes is a good default.

What refreshing means: not only ingesting new data (spends for past days, week) but also the model updates itself (“learns on new data”).

For using MMM you don’t need to understand statistics – MMM results should be easy to interpret by business users (if they are not, it is a sign of a low-quality model and/or its vendor).

For developing MMM you (or your team or external vendor) need to understand statistical learning but also – and this is sometimes underestimated by general data scientists trying to develop MMMs on their own from scratch – experience.

You don’t. What your organization needs is some maturity in working with data and measurement – e.g. understanding the concept of incrementality or being comfortable with working with uncertainty. You should also see that there is a need for marketing measurement and how it can impact resource and budget decisions in your organization. 

MMM results are of course often used by analysts but also (or even primarily) by senior people in the marketing department and other departments like finance who are responsible for budgeting, planning and revenue delivery.

In principle yes and some companies do it – esp.those with strong inhouse data science teams and previous experience with MMM. There are even some open-source libraries that can help you – however beware that there is a huge difference between “getting a first-attempt model using some open-source library” and “production-useful MMM” – there is a high chance that your first models will be useless at best, highly dangerous for your business at worst.

Another aspect that we often see underestimated by inhouse attempts at MMM is continuous development and support: to actually get the benefits of using MMM, you should think of MMM as an iterative process – you develop a model, you use it, gather feedback, get new calibration points, then you improve the model etc. so if you are thinking about inhouse MMM development, you should commit the data science/data resources for the long term, not just for initial development.

MMM - What it can and cannot do

Generally all marketing channels as long as you have a time series (ideally daily or weekly) of their costs and/or other measure of their volume – impressions, GRPs etc.

So for example Google Ads (often split into multiple sub-channels based on campaign or objective type), FB/Meta Ads (also often split into multiple smaller channels), TikTok, Amazon Ads and retail/commerce media in general, Linkedin, Twitter/X ads, Pinterest, Snapchat, display ads, video ads, influencer marketing, PR activities, events, TV/CTV, OOH, radio, print advertising, leaflets are examples of channels that are usually

There are some cases where MMM may not be able to accurately measure channel performance so keep these in mind

    • Small channels – if the spend share of a channel is less than 2% it is often difficult to distinguish its effect from overall noise
    • No or very low variance – if there is no variance in the intensity of the channel, this is a problem. These could be sponsorships that run for the whole investigated period of time (e.g. last 2-3 years) with no change.
    • Affiliate marketing – because of the mechanics of how it works, this presents a specific challenge in many cases
    • Completely new channels – esp when there are no similar channels in your recent history. This is usually solvable just keep in mind that measuring completely new channels is a bit challenging in MMM.
    • Channels that are always run together with the same intensity – eg if you run Youtube and TV ads always at the same dates and with same relative intensity, it may not be possible to distinguish between these two channels and only the aggregate effect of both would be identified

Existing experiment results can significantly speed up any MMM program and will almost certainly lead to a faster value and ROI on your MMM project.

Set up a good data collection process for areas like promotions (promotion calendars and plans), main events affecting your business, discounting or major pricing changes etc.

MMM can be used to e.g. understand how Youtube or TV ads impact your sales on Amazon, Walmart or other marketplaces. Or how it impacts your sales with retail partners or in your own retail store network – this is one of the very frequent use cases of MMM. 

  • So let’s say you are a DTC brand with the following sales channels
    • Own website
    • Amazon marketplace
    • Retail partners

And you run a campaign on Instagram, Youtube, TikTok and CTV – using MMM you can see the overall ROI of these medias for each of the sales channels.

Branding and long-term mean different things for different companies. Generally MMM is quite reliable if the results (sales effect) comes up to 3-4 months after the activity. For this use case standard / modern MMM techniques are suitable and good enough.

Measuring accurately truly long-term effects of marketing where you communicate to potential customers who are not in-market and may not be for many months (and you are trying to mostly build mental availability) is difficult and standard MMM is not suitable for this. There are ways to extend MMM to cover even these use cases – this will inevitably require a much longer time series of data (think 4-5 years instead of 2 years) and additional techniques.

MMM - Specific media and channels

They are not really the focus of MMM but in some cases it may be important for the advertiser to have some insights into them even via MMM – let’s discuss them one by one

  • Emailing (or similar – push notifications etc) – if you have some good metric representing the volume of the activity (e.g. number of emails opened – which would be somewhat analogous to impressions), then it is generally possible to include it in MMM.
  • Direct is a complicated “channel” nowadays – to some extent it represent navigational “brand” traffic but in recent years more and more of “direct” is actually tracking issues or more generally unknown source of traffic, so in reality “direct” that you see in your web analytics tool is a mix of these two. In MMM 
    • Revenue driven by your brand strength has its own “bucket” called brand baseline
    • MMM is not affected by any tracking issues, so if the revenue (or other modelled outcome) is caused by a channel, eg Facebook, it will be attributed to Facebook – irrespective of any tracking issues.
  • Organic (from search engines) can be modelled using various proxy metrics (e.g. number of organic impressions from GSC) but often adds a lot of complexity to the model so unless there is a specific case to be made to measure organic this way, we recommend leaving it as one component of the baseline revenue.

Yes these are the typical use cases where MMM is very strong and offers a better solution than most (even data driven) attribution models. As MMM does not need any tracking it is not affected by typical tracking issues or the fact that many formats on these platforms work (influence consumers) without generating clicks.

Yes – this is another good use case for MMM. It requires some specific data preparation but a good MMM solution should have no problem with influencer marketing.

Yes – another typical use case (and totally ignored by attribution models). There are multiple ways to integrate price changes and/or discounts into MMM – each of them having pros and cons. One thing you need to do: collect the necessary data for these in a time series – unless you do so already.

MMM - For whom is it actually suitable?

MMM used to be the territory of Fortune 500 brands but with advances over the past cca 5 years, it has become much more accessible. But it is still quite a complex undertaking. Our benchmark is that if your brand spends at least 2 million USD/EUR annually in media (online or offline), MMM should be relevant for you and the insights and optimization should easily offset the complexity and cost of MMM.

For smaller brands it is still possible to implement an MMM solution – there is no hard minimum limit on marketing spend – but it may require a more careful consideration whether it makes sense in terms of your focus and potential business impact.

MMM is relevant for many B2B categories too – e.g. many B2B technology and SaaS companies use it. Depending on the length of your sales cycle you may want to either model actual sales or the number of MQLs or SQLs as the modelled KPI.

There are some industries with very long sales cycles and too complex distribution channel systems that may not be suitable – it is best to discuss your specific case with your MMM partner.

Most brands can benefit from MMM but there are some cases where in our experience either delivering a reliable MMM is often challenging or the business impact may not be material:

    • New brands or products – MMM needs some history (realistic minimum is 1 year of marketing data, 2 years are recommended)
    • Smaller businesses – MMM is still a quite complex undertaking and if you only spend maybe 50-100k USD on marketing a month, you may see a better ROI on other initiatives than advanced analytics as MMM. For smaller brands we recommend to start using incrementality testing – this is the best way to build solid foundations for future MMM projects.
    • Brands with very long sales cycles – eg. B2B industrial equipment etc.
    • Rapidly changing or nascent industries – while MMM can handle changing market conditions and trends, if your industry is undergoing a complete disruption, it may be difficult to form a reliable MMM.
    • Very complex distribution channels – some companies may struggle to track end sales in time accurately e.g. if they sell through a multilayered system of distribution partners.

MMM - Why should we trust it? Model results validation

There are some standard statistical measures than can be used to assess the model quality, most often used are:

  • R-squared – this measure says to what extent the model can explain variance in the modelled KPI. General rule of thumb is
  • R squared > 0.9 = great result
  • R squared between 0.8 – 0.9 = good enough for production use
  • R squared below 0.8 = model probably needs some changes (eg different channels need to be introduced)
  • It is also necessary to distinguish between R-squared on training data (=the model has seen both independent and dependent variables in the data set, so typically daily/weekly channel spends and total sales) vs R-squared on test (or hold-out data) – where the model has not seen the real values of the dependent variable and needs to predict it. This is a much more difficult task and the R-squared here will be lower.
  • We recommend looking at MAPE (Mean Absolute Percentage Error) which expresses the error as a percentage of the actual values.

Besides that the best way to develop trust in the model results is to perform tests and compare their actual results with what the model predicts – more info is in the section on model validation

In general there are 2 main ways:

  • Experiments – here you basically make a change (e.g. increase or decrease spend in a specific channel) for some period, then you evaluate the effect using some independent method (causal inference) and compare the result with what MMM is telling you. If the results are close, it indicates that the model can make good predictions, if it is not, you should use the experiment result as an additional calibration point to correct the model.
  • Hold-out data testing – this is a general best practice in statistical modelling. The principle is that you don’t show the model all the data – you “hide” e.g. last 2 months. So you know what actually happened in those last 2 months (how much you invested in each channel, what were the total sales) but the model doesn’t know this data. Then you show the model the channel spends for the last 2 months (but not the sales) and you let it predict what it thinks the total sales would be. Then you compare the actual sales (you know them) with model prediction.